An HMM-based method has been developed for the tone recognition of trisyllabic speech of standard Chinese. Several schemes developed for the tone recognition of dissyllabic speech were also incorporated: use of macroscopic and microscopic parameters of fundamental frequency contours, speaker normalization with fundamental frequency offset, coping with tone sandhi effect by additional tone models, concatenated learning of HMM, etc. Adding to these, a scheme of double codebooks on fundamental frequency and power was newly incorporated to the method to cope with the rather low recognition rate for light tones. By further discriminating light tones from third tones using syllabic durations, the recognition rate of 75% was obtained for the light tone. The total recognition rate was reached 97.2%.